There is a number buried in a new survey that should stop any founder shipping agents in production. To resolve a single deep research problem, an agent may call a search API roughly six hundred times. Not six. Six hundred. Each of those calls drags a payload of tokens back into the context window, and every one of those tokens gets re-read on the next turn, and the turn after that, and the turn after that.
The paper is called Toward Efficient Agents: A Survey of Memory, Tool Use, and Planning, by seventeen researchers spanning nine institutions. It runs to 63 pages and 244 references, and it tabulates 117 distinct methods. What makes it worth your attention has nothing to do with its size. The survey asks a question the industry has been quietly avoiding while it celebrates capability: when an agent gets smarter, who pays, and how much?
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.